Agentic Craft

Memory

Patterns for persistent memory, cross-session recall, context awareness, auto-detection of preferences, and privacy controls for reviewer data.

Persistence

Memory Panel

The reviewer's stored preferences and facts — key-value pairs the agent uses to personalize responses across sessions.

Controls

Memories6

enterprise release

Workspace

Preferred release tier

Launch review session12m agoExpires 90 days

Launch Policy v2

Project

Default launch policy

Project brief v331m agoExpires 180 days

Launch review team

Team

Review workspace

Workspace settings1h agoExpires Never

Markdown with operating playbook work unit headers

User

Report format

User preferenceYesterdayExpires 90 days

self-serve onboarding flow

Project

product type

Product notesYesterdayExpires 60 days

Internal launch process

Team

Approval team

Approval policy2d agoExpires 180 days

State	Content	Behavior
Empty	Placeholder with guidance text	Shown when the agent has no stored memories
Populated	Key-value list with hover actions	Edit and delete icons appear on hover per row
Search	Filtered list with search input	Real-time filtering by key or value

Memory entries are key-value pairs rather than free text — this makes them scannable and editable without requiring the reviewer to parse unstructured prose. Actions appear only on hover to keep the list clean.

Management

Memory Entry CRUD

View, edit, and delete individual memory entries. Each operation uses inline controls so the reviewer stays in context.

Controls

Preferred release tier

enterprise release

Operation	Interaction	Behavior
View	Default display	Shows key-value pair with edit and delete icons
Edit	Click edit icon	Inline text field replaces value, Save/Cancel buttons appear
Delete	Click delete icon	Confirmation prompt with warning text and Confirm/Cancel

All CRUD operations happen inline — no modals. The delete confirmation is intentionally low-drama: a text warning, not a blocking dialog. This matches the lightweight feel of key-value memory management.

Detection

Auto-Memory

The agent detects learnable information from conversation and offers to save it as a memory — the reviewer always has the final say.

Controls

I see you've been consistently requesting enterprise release reviews across the last three sessions. The reference document Launch Policy v2 appears to be your default reference.

I noticed you prefer enterprise release — save this preference?

State	Visual	Behavior
Detected	Banner below agent message with Save/Dismiss	Agent surfaces the detected preference with a clear opt-in prompt
Saved	Brief confirmation message	Fades in, replacing the banner. Memory is persisted immediately
Dismissed	Banner removed, muted note	Agent will not re-prompt for this particular preference

Auto-memory is opt-in by design. The agent detects patterns but never persists without explicit reviewer consent. This is critical for regulated environments where auditors may need to justify what data the tool retains.

Awareness

Memory Context Ring

Visual indicator showing how much memory context is loaded into the current session. Details are revealed on hover to avoid visual clutter.

Controls

No memory loaded

Start a conversation to load reviewer preferences and context.

State	Ring Visual	Tooltip
No Context	Empty ring with muted icon	No tooltip — nothing to show
With Context	Partially filled ring, animated on load	Lists loaded memories with key-value pairs on hover or focus

No number inside the ring — the icon signals presence, and the tooltip provides detail on demand. This avoids cognitive load for reviewers who don't need to know the exact count at a glance.

Governance

Privacy Controls

Category-level toggles for what the agent is allowed to remember. Reviewers working under NDA or handling sensitive product data may need to disable specific categories.

Controls

Preferences

release tier, report format, workflow choices

Document History

Previously reviewed briefs, policies, and launch summaries

Project Context

Current product, team, and approval path

Personal Info

Name, timezone, role at review team

Preset	Toggle State	Notes
All On	All categories enabled	Default for most reviewers
Selective	Preferences + Documents on, Context + Personal off	Common for consultants working across multiple review teams
All Off	All categories disabled with confirmation	For sensitive reviews or private review sessions

Privacy controls are category-level, not per-entry — this keeps the mental model simple. Disabling a category does not delete existing memories, it just prevents the agent from reading or updating them. This distinction matters for complex review workflows where data retention policies may differ from data access policies.